Deskilling

What it is

Krzysztof Budzyń and colleagues studied four endoscopy centres in Poland that introduced AI tools for polyp detection at the end of 2021, as part of the ACCEPT trial. After the tools arrived, colonoscopies were assigned to be done with or without AI depending on the date. The researchers compared standard colonoscopies, done without AI, in the three months before and the three months after the AI was introduced. Across 1,443 patients, the adenoma detection rate in non-AI colonoscopies fell from 28.4 per cent to 22.4 per cent. Exposure to AI was independently linked to the lower rate, along with the patient's sex and age. The study was published in The Lancet Gastroenterology & Hepatology in 2025 and was widely reported.

In effect

Deskilling matters for any service that relies on AI but cannot guarantee it will always be running, and for any plan that relies on clinicians to check an AI's work. A clinician whose own detection has slipped is less able to catch the machine's misses. It sits alongside automation bias, measured by Thomas Dratsch and colleagues in 2023, in which radiologists followed wrong AI suggestions on mammograms.

What it does not say

The colonoscopy study was observational. It compared two periods rather than randomly assigning doctors to use AI or not, so other changes over the same months could account for part of the fall. It does not show that AI-assisted colonoscopy is worse overall: randomised trials find that detection rises when the AI is in use. The question it raises is about performance without the tool, and how quickly skills return or decline, which needs further study.

---